Training Compute Thresholds: Features and Functions in AI Governance

This paper examines the use of training compute thresholds as a tool for governing AI systems. We argue that compute thresholds serve as a valuable trigger for further evaluation of AI models, rather than being the sole determinant of the regulation. Key advantages of compute thresholds include their correlation with model capabilities and risks, quantifiability, ease of measurement, robustness to circumvention, knowability before model development and deployment, potential for external verification, and targeted scope. Compute thresholds provide a practical starting point for identifying potentially high-risk models and can be used as an initial filter in AI governance frameworks alongside other sector-specific regulations and broader governance measures.